Research map / Label-efficient and robust learning / Recognition and long-tail
BD-Merging: Bias-Aware Dynamic Model Merging with Evidence-Guided Contrastive Learning
Yuhan Xie, Chen Lyu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add debiased router (per-sample merging weights) (ablation) | BD-Merging · 8 tasks, corrupted L2 | avg accuracy: 67.25 -> 78.78 | Improved |
| Add debiased router (per-sample merging weights) (ablation) | BD-Merging · 8 tasks, clean | avg accuracy: 78.31 -> 87.15 | Improved |
| Add discrepancy-aware contrastive loss L_Dis (ablation) | BD-Merging · 8 tasks, corrupted L2 | avg accuracy: 74.26 -> 78.78 | Improved |
| Add Adjacency Discrepancy Score (ADS) (ablation) | BD-Merging · 8 tasks, corrupted L2 | avg accuracy: 75.44 -> 78.78 | Improved |
| Add semantic divergence term Div within ADS (ablation) | BD-Merging · 8 tasks, corrupted L2 | avg accuracy: 76.28 -> 78.78 | Improved |
| Add opinion conflict term Conf within ADS (ablation) | BD-Merging · 8 tasks, corrupted L2 | avg accuracy: 76.95 -> 78.78 | Improved |
| Add prediction sharpness term Sharp within ADS (ablation) | BD-Merging · 8 tasks, clean | avg accuracy: 86.68 -> 87.15 | Improved |
| Add inverse correlation loss L_Inv (ablation) | BD-Merging · 8 tasks, corrupted L2 | avg accuracy: 77.43 -> 78.78 | Improved |
| Use BD-Merging instead of Twin-Merging on unseen tasks | CLIP ViT-B/32 merged model · 4 unseen tasks | avg accuracy: 53.03 -> 55.01 | Improved |
| Merge models (BD-Merging) vs. pretrained CLIP on unseen tasks | CLIP ViT-B/32 · 4 unseen tasks | avg accuracy: 56.99 -> 55.01 | Hurt |
| Increase corruption severity from L1 to L3 | BD-Merging (layer-wise) · 8 tasks | avg accuracy: 83.31 -> 75.36 | Hurt |
| Use BD-Merging layer-wise instead of AdaMerging w/ Surgery under corruption L2 | CLIP ViT-B/32 · 8 tasks, corrupted L2 | avg accuracy: 74.33 -> 78.78 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdaMerging — AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.87)
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.87)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 2025 · similarity 0.87)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.87)
- CALM: Consensus-Aware Localized Merging for Multi-Task Learning (ICML 2025 · similarity 0.86)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
